Sabitlenmiş Tweet
I just released a new feature for Hyperfolio and the Hyperfolio API: Yield.
You can now browse thousands of yield opportunities on HyperEVM across nearly 30 protocols
English
stableAPY.hl
2.5K posts
stableAPY.hl
@stableAPY
Building HyperFolio - HyperEVM Portfolio Tracker
Katılım Şubat 2025
257 Takip Edilen2.1K Takipçiler
@stableAPY will plug a 3060 into my mac mini 24GB soon
just for lulz
was planning to use 3060 as my first perma worker node
acemagic 128GB incoming soon
English
stableAPY.hl retweetledi
I still can't get my head around the fact that my 3060 12gb is running Qwen 3.6 35B at 40 tok/s
this card is around 200$ second-hand
while everyone's shilling super expensive 128gb unified memory or RTX 6000 cards, one 3060 12gb can be far from enough for first local AI experiments
it's cheap, and paired with a bit of RAM and a somehow decent CPU it'll do the work
of course decode drops when context grows and you won't be able to run multiple sub agents, but it's a nice cheap entry point
for example, it pairs very well with my 3090:
> 3090 running main agent 35B -np 2 => so I can have 2 concurrent agent
> 3060 sub-agent 35B -np1
this way my main hermes can delegate work to this sub agent while working on something else
I also run Hermes cron job so they does not overload the main agent and I don't mind it behind slower because it happens in the background
English
@stableAPY @ItsmeAjayKV @UnslothAI Try the 4 k M version, it seems to be much faster. Have consistently seen IQ cause slower dequant even at smaller model sizes (6 K M also good too)
English
it seems that Unsloth is all in on MTP (Multi-Token Prediction)
they released Qwen 3.5 0.8B, 2B, 4B and 9B MTP versions
looks like I'll have new stuff to test out on my 3060
huggingface.co/unsloth/Qwen3.…
English
@Russian_ikigai not the MTP on this screen but had some good tests today for MTP on the 3060 might use it as the "prod" setup
English
@jonorozcoc I find it very good!
i've also tried gemma, but not much real coding test, more benchmarks : x.com/stableAPY/stat…
stableAPY.hl@stableAPY
After testing out Qwen 3.6 35B on my 3060, it's Gemma 4 26B A4B UD-IQ4_XS's turn sadly it seems that it requires more CPU offloading than Qwen, leading to lower prefill and decode speeds while it felt counterintuitive, it's actually pretty logical even though Gemma is 26B vs 35B for Qwen, the 4B active parameters lead to higher CPU offload I'll still try it out to see how good the model is. I've been hyperfocused on Qwen and it's time to see what other labs have to offer
English
@darth_turnip try ik_llama.cpp as a start, then try their MTP maybe
x.com/stableAPY/stat…
stableAPY.hl@stableAPY
small GPU owners, 3060's fellows, 8 or 12gb VRAM chads, > if you're using llama.cpp you should check out ik_llama.cpp it gives about 10% better performances at 128k total context running Qwen3.6-35B-A3B IQ4_XS on a 3060 12gb 31,7 tok/s at 50% of context is not that bad for a subagent for local hermes setup github.com/ikawrakow/ik_l…
English
@stableAPY 12GB VRAM. Various back ends, Ollama, Kobold. I have 64GB of system RAM.
English
@darth_turnip I think behind Q3_XL it's pretty degraded
I use IQ4_XS, what version of the 3060, 8 or 12gb? What backend?
kaitchup.substack.com/p/summary-of-q…
English
@stableAPY Qwen3.6-35B-A3B-UD-IQ2_M - What would you recommend?
English
@BitcoinComfy @georgecursor bought mine 750 eur and got another one for around 600, there are some opportunities, but I've alerts and stuff
English
@stableAPY @georgecursor 3090 is around 900/950 (was half price 1y ago)
English
@stableAPY Are they still asking us to load a special fork of llama.cpp?
English
@stableAPY Mine was kinda dumb 😂
Have to test it more, just started yesterday.
English
stableAPY.hl retweetledi
coinbase announcement summary:
> expected to add ~$150M in annual revenue (~25% increase)
> coinbase will use builder codes to relaunch perps on its platform
despite the news, hyperliquid:native is only up 4% today
English
stableAPY.hl retweetledi
The one big thing I'm waiting for in llama.cpp (shout out to @ggerganov ) right now is MTP. With Qwen 3.6 (my fav model) already supporting it, we are going to see massive improvements in generation speed once it's fully merged.
So, what exactly is MTP?
It stands for Multi-Token Prediction. If you understand speculative decoding, this is the next level. Instead of relying on a smaller, separate draft model, MTP is built directly into the model during its initial training. The main model simply produces draft tokens on its own auxiliary heads that allow it to naturally output multiple future tokens simultaneously. It's leaner, faster, and incredibly efficient for local hardware.
How is it different from other methods. Well lets go over them in brief.
1. Standard Speculative Decoding (Draft models)
You load two models into memory: the big target model (e.g 35B) and a tiny fast draft model ( < 2B) from the same family. The small draft model runs ahead, generating 4 or 5 tokens sequentially. The massive target model then does a single forward pass to check the drafts math.
Pros: Consistent speedups across workloads.
Cons: Eats more VRAM, if i talk about my 3060 case, where i try to squeeze a heavy model into 12GB of VRAM, sacrificing a GB or two just to host a draft model can be a painful trade-off.
2. n-gram speculative decoding (prompt lookup)
This one is interesting, different idea than draft, n-gram decoding simply looks at the text already in the "prompt" and guesses that it will be repeated (which is also its biggest issue). Good for coding, JSON formatting, or even rag.
Pros: Zero VRAM overhead. Nothing extra to load. Good speedup for above mentioned tasks.
Cons: Very situational. For creative writing it fails miserably and offers almost no speedup.
3. DFlash (Block diffusion drafting)
DFlash replaces traditional autoregressive draft model with a lightweight block diffusion model. Instead of guessing tokens sequentially, DFlash generates an entire block of tokens in parallel in a single forward pass. It achieves this by pulling hidden state features directrly from the target model andusing them as context to denoise a block of next tokens immedietly.
Pros: Super fast, by removing sequential bottleneck of drafting phase, this can achieve high loseless acceleration.
Cons: Nothing much actually, it does requires specialized checkpoints trained specifically to align with the target model.
Also take a look at LuceBox-hub D-Flash and P-flash by @davideciffa
English
@0xasrequired "5) coinbase gives up on fighting hyperliquid and kisses the ring, as required."
> my favorite part
English
stableAPY.hl retweetledi
1) what
2) it’s joever for cash and hyna, unless they cough up 90% of USDT0/USDe’s cost adjusted rev and even then, the fragmentation will get them
3) can’t imagine anyone launching a HIP-3/4 DEX that doesn’t use USDC anymore
4) no more fragmentation: cross margin between validator operated perps, xyz, spot, HIP-4, km, flx, vntls and para, as required.
5) coinbase gives up on fighting hyperliquid and kisses the ring, as required.
6) @fiege_max conquered goliath
7) i thought circle already cut a huge rev share deal with lighter to keep them from making their own stablecoin, a la USDH
8) coinbase has to buy and stake 500k HYPE
9) AQAv2 no longer has the 20% taker fee discount, 50% better maker rebates and 20% more vol contribution to fee tiers
10) hyperliquid won forever, as required.
Hyperliquid@HyperliquidX
Coinbase has announced its plan to activate AQAv2 on USDC as the treasury deployer, with Circle serving as the technical deployer responsible for CCTP and native cross-chain infrastructure. Both Coinbase and Circle have committed to stake HYPE to activate AQAv2. As part of this transition, Native Markets has agreed to terms granting Coinbase the right to purchase the USDH brand assets. With Coinbase, in its role as treasury deployer, sharing the vast majority of reserve yield revenue with the protocol, USDC will become the most aligned stablecoin on Hyperliquid. As a result, canonical outcome (HIP-4) markets will use USDC as the quote asset in a future network upgrade. User and builder feedback has been consistent that fragmentation leads to degraded experience; now, the community no longer needs to choose between liquidity and protocol alignment. The pioneering work of Native Markets in launching USDH as the first production-scale stablecoin sharing yield directly with a protocol in a purely onchain implementation made AQAv2 possible. The learnings and mechanics pioneered by USDH will live on in AQAv2. The Hyper Foundation will give grants to eligible HIP-3 deployers, HIP-1 deployers, and builders who integrated USDH, supporting teams through migration over the next months. These grants reflect an ongoing commitment to teams who choose to build on Hyperliquid and align with the protocol. USDH markets are fully functional but will sunset over time. USDH remains fully backed, with feeless conversions to USDC and fiat available to users during this transition.
English